Loose posts

A new Microsoft and City University of Hong Kong paper calls an agent's ability to pick the right next step on long tasks "taste" and built a 502-fork decision test set, on which the strongest model scored only 59.7%.

The creator had the model score each part of a scene and keep improving it until it reached at least 8 out of 10, saying Opus 5.5's detail and realism exceed other models.

Tencent chief AI scientist Yao Shunyu now also heads the AI Data department, putting the company's models, infrastructure, data and evaluation under his unified management.

The author argues the real takeaway is the process of large-scale parallel exploration, anomaly detection and human verification, not the discovery itself.

A new method from Google and others uses an optimized harness only during training, raising macro task success from 23.3% to 44.3% without a specialized harness at deployment.

According to The Information, DeepSeek's revenue doubled mainly through price increases, with a planned STAR Market raise of about $7.45B, while Liang Wenfeng says over 70% of compute still goes to training.

Claude identified repeated DNA sequences next to a phage gene, which the team named ART; the system resembles CRISPR in structure but its function is unknown.

Boltzbit and a Cambridge team use a hypernetwork to dynamically compile new facts from conversations into LoRA weights, beating long-context methods on long documents and multi-turn dialogue with constant per-turn compute.

About 950 Claude agents ran for 21 hours and screened over 200,000 reverse transcriptases to find a new system, which Feng Zhang called a good example of AI participating in biological discovery.

Given 13 takes of a flubbed line, the model read the transcripts, compared the top three frame by frame, picked the best, cut, graded and captioned it, rewrote its own captions, and spent 90 cents in 2 minutes.

Without any AI video tool, the model wrote the storyboard, drew each frame in p5.js, screenshotted them in headless Chrome, stitched them with ffmpeg and added the song, assigning parallel sub-agents to edit each chapter.

Anthropic's bio lab published its first results: about 950 Claude agents ran for 21 hours and screened DNA databases down to 3,500 new candidate reverse transcriptase systems, named ART.